tinker_cookbook.preference.PreferenceModelFromChatRenderer
class tinker_cookbook.preference.PreferenceModelFromChatRenderer(PreferenceModel)
A PreferenceModel that uses a chat renderer and Tinker sampling client.
Renders a Comparison into a prompt, samples a single token (A, B, or Tie), and maps it to a float score (-1 for A, 1 for B, 0 for Tie).
__call__(comparison)
Score a comparison by sampling a preference token from the model.
Renders the comparison into a prompt, samples a single token
(A, B, or Tie), and maps it to a float.
Parameters:
- comparison (Comparison) – The comparison to evaluate.
Returns: float – -1.0 if A is preferred, 1.0 if B is preferred, 0.0 for a tie or an unrecognized output.